Abstract: Detecting embryo viability is essential to the quality and safety of vaccine production, especially in large-scale manufacturing. Rapid and accurate detection of embryo viability can improve the production efficiency for the final quality of vaccines. Traditional machine vision detection can rely heavily on the complex algorithms of feature extraction, most of which are often designed for specific scenarios. However, the detection accuracy and stability are also sensitive to the image quality and environmental conditions, such as lighting, background, or temperature. Additionally, the applicability of traditional detection has been limited to fault tolerance in different environments, when dealing with noise or abnormal conditions. To address these challenges, this study aims to detect the vaccine embryo viability using an improved YOLOv8 model. Several innovations were incorporated to enhance efficiency, accuracy, and adaptability. A specialized system of image acquisition was developed to capture the high-quality images of embryos incubated for 10 to 11 days. The consistent dataset was obtained in the varying environmental conditions. The dataset was then expanded using geometric transformations, color adjustments, and image enhancement. As such, the robustness of the model increased to handle the diverse image conditions. In terms of model improvements, ShuffleNetV2 was used to replace the YOLOv8 backbone. Computational complexity was significantly reduced to maintain high accuracy, indicating more suitable for deployment on embedded devices where computational power was limited. The overall efficiency of the model was enhanced to support its application in large-scale industrial environments. Additionally, a dynamic snake convolutional layer was added to the neck of the YOLOv8 model. This layer was used to adaptively focus on the elongated and curved structures in embryos, in order to capture the geometric features of tubular structures. The precision of detection was improved to more accurately assess the physiological state of the embryos. Furthermore, the EIoU (Embedding Intersection over Union) loss function was introduced to more effectively detect the boundary box alignment and shape similarity, compared with the traditional IOU. EIoU improved the accuracy of boundary box positioning, while reducing the errors related to the complex shapes of embryos, thereby enhancing the reliability of the model in real-world applications. Experimental results confirmed that the superior performance of the improved YOLOv8 model was achieved to detect embryo viability. There was a precision of 99.2%, a recall of 98.2%, and a mean average precision (mAP_50-95) of 96.9%, with increases of 2, 0.3, and 1.5 percentage points, respectively, compared with the original YOLOv8 model. Additionally, the computational complexity and inference time were reduced by 60.9% and 60.5%, respectively. The improved model was highly suited for the large-scale detection of embryos. The finding can also provide an efficient, non-destructive approach for the rapid detection of the vaccine embryo viability.